# Americas priority claim ledger

**Cut-off:** 2026-08-11 (Europe/Madrid)
**Status rule:** an observed item in an unfinished election is coded `ongoing`, never as a demonstrated electoral effect. Reach is not treated as behaviour, and behaviour is not treated as vote change.

## Evidence shorthand

- `established_evidence`: directly supported by a final administrative/court record, an actor admission, or convergent authoritative records.
- `strong_inference`: the facts support the stated mechanism, but the mechanism itself was not experimentally observed.
- `plausible_hypothesis`: a material link remains unresolved.
- `open_question`: the available evidence does not answer the claim.
- Confidence applies to the claim **as worded**, not to the overall case.

---

## 1. `usa-2024-nh-biden-ai-robocall`

### Proposed case record

| Field | Value |
|---|---|
| `case_id` | `usa-2024-nh-biden-ai-robocall` |
| `case_family_id` | `usa-2024-presidential-election` |
| Country | United States (New Hampshire) |
| Election/date | New Hampshire Democratic presidential primary; calls on 2024-01-21; primary on 2024-01-23 |
| Incident | AI-cloned Biden “save your vote” robocall |
| Status | `retrospective` |
| Actor / attribution | Political consultant Steve Kramer: `admitted` as orchestrator and `adjudicated` by the FCC for the caller-ID violations. Voice Broadcasting/Life transmitted the campaign; Lingo originated 3,978 completed calls and resolved the FCC investigation by consent decree. Criminal intent must retain the separate 2025 acquittal. |
| AI/digital capability | Confirmed ElevenLabs voice clone; voter-list selection; automated telecom delivery; caller-ID spoofing; A-level STIR/SHAKEN attestation by the originating carrier |
| Influence vector/platform | Telephone robocall network; no social-platform virality is required for the mechanism |
| Target | Potential New Hampshire voters; the supplied list was labelled as prior Democratic-primary voters and Kramer said he sought likely Democrats |
| Mechanism | A synthetic presidential voice falsely implied that a primary ballot would reduce or consume the recipient's November vote, told recipients to “save” the vote, and displayed the number of a local Biden write-in organizer as caller ID |
| Scale/reach | 9,581 call attempts initiated; 3,978 calls completed by Lingo. These are delivery records, **not 9,581 unique listeners or persuaded voters**. |
| Behavioural/electoral effect | Unknown. No credible study or official record measures abstention, turnout change, candidate-choice change, or a changed result attributable to the calls. Biden's write-in win does not prove either effect or no effect. |
| Agency-transfer path | Voice cloning + voter-list targeting + spoofed local provenance transferred control over the apparent source and urgency of a participation instruction from the voter/Biden/local organizer to Kramer. The recipient's capacity at risk was informed, autonomous participation. |
| Democratic harm | Attempted/deceptive voter discouragement; impersonation; corruption of caller-authentication signals; burden on the impersonated organizer, regulators, civic groups, and voters to correct the message just before voting |
| Response/outcome | FCC final $6,000,000 forfeiture against Kramer; Lingo consent decree with $1,000,000 civil penalty and compliance plan; New Hampshire prosecution ended in acquittal on all 22 counts in 2025; a separate federal civil case ended in default judgment, monetary relief, and a nationwide injunction against Kramer. |
| Evidence classification | Occurrence: established; call delivery: platform/telecom measured; attribution: admitted + final FCC finding; administrative legal outcome: final; criminal liability: acquitted; behavioural effect: unknown; electoral effect: unknown |
| Overall confidence | 0.96 for occurrence/mechanism/actor; 0.99 for call-attempt/completion counts; 0.15 for any causal electoral-effect proposition |

### Atomic claim ledger

| Claim ID | Type | Atomic claim | Label | Conf. | Supporting sources | Counterevidence / constraint |
|---|---|---|---|---:|---|---|
| `clm-nh-001` | occurrence / ai_role | On 2024-01-21 an artificial voice made to sound like President Biden was delivered in robocalls to potential New Hampshire voters two days before the primary. | established_evidence | 0.99 | `src-nh-fcc-final`; `src-nh-fcc-lingo` | None material on occurrence. |
| `clm-nh-002` | mechanism | The message falsely linked primary participation to the ability/value of voting in November and urged recipients to “save” their vote; the caller ID used a number associated with an uninvolved local Biden write-in organizer. | established_evidence | 0.99 | `src-nh-fcc-final`; `src-nh-lwv-case` | Kramer later characterized the operation as an AI “wake-up call”; that account concerns intent, not the content or spoofed number. |
| `clm-nh-003` | attribution | Kramer supplied the script, audio, voter list, timing, and spoofed number and later publicly accepted responsibility. | established_evidence | 0.99 | `src-nh-fcc-final`; `src-nh-ap-acquittal` | His client, the Dean Phillips campaign, denied authorizing the operation. Do not attribute the call to Phillips or his campaign. |
| `clm-nh-004` | reach | Voice Broadcasting initiated 9,581 calls; Lingo's records show 3,978 completed calls bearing A-level attestations. | established_evidence | 0.99 | `src-nh-fcc-final`; `src-nh-fcc-lingo` | Attempted calls are not completed calls; completed calls are not verified human listens; neither number is unique exposure. |
| `clm-nh-005` | intent / legal_status | The FCC made a final administrative finding that Kramer knowingly caused misleading caller ID with intent to defraud and cause harm and imposed a $6 million forfeiture. | established_evidence | 0.99 | `src-nh-fcc-final` | In a different proceeding with different statutes and burden of proof, a New Hampshire jury acquitted him on 11 voter-suppression and 11 candidate-impersonation counts. Do not translate the FCC finding into a criminal conviction. |
| `clm-nh-006` | institutional_outcome | The originating carrier, Lingo, agreed to a $1 million civil penalty and a compliance plan after the FCC found it completed 3,978 calls with A-level attestations. | established_evidence | 0.99 | `src-nh-fcc-lingo` | A consent decree resolves the FCC investigation; it is not a criminal adjudication. |
| `clm-nh-007` | legal_status / counterevidence | On 2025-06-13 a jury acquitted Kramer on all 22 New Hampshire criminal counts. | established_evidence | 0.99 | `src-nh-ap-acquittal` | The acquittal does not erase the final FCC forfeiture, the admitted orchestration, or the later civil default judgment. |
| `clm-nh-008` | institutional_outcome | In November 2025 a federal court entered default judgment against Kramer, awarded relief to the individual plaintiffs, and permanently enjoined similar voter-coercion schemes. | established_evidence | 0.95 | `src-nh-lwv-case`; `src-nh-ap-default` | Default judgment followed Kramer's nonappearance; describe it as a default, not as a contested merits trial. |
| `clm-nh-009` | behavioural_effect | The public record does not establish how many recipients believed the message or abstained because of it. | open_question | 0.98 | `src-nh-fcc-final`; `src-nh-ap-acquittal` | Call counts and media coverage cannot fill this gap. |
| `clm-nh-010` | electoral_effect | No evidence establishes that the calls changed the primary result or any later election result. | open_question | 0.98 | `src-nh-ap-acquittal`; `src-nh-lwv-case` | Biden won the write-in campaign; this is contextual counterevidence to dramatic outcome claims, not a causal estimate of zero effect. |
| `clm-nh-011` | agency_transfer | The combination of a cloned presidential voice, a politically selected call list, and a spoofed trusted in-state number plausibly increased the actor's control over the perceived authorship and relevance of an abstention cue. | strong_inference | 0.90 | `src-nh-fcc-final`; `src-nh-fcc-lingo` | The mechanism is strongly evidenced; its effect on any individual recipient is unmeasured. |

### Sources

| Source ID | Role / quality | Exact URL |
|---|---|---|
| `src-nh-fcc-final` | FCC final forfeiture order; primary record; A | https://docs.fcc.gov/public/attachments/FCC-24-104A1.pdf |
| `src-nh-fcc-lingo` | FCC Lingo order and consent decree; primary record; A | https://docs.fcc.gov/public/attachments/DA-24-790A1.txt |
| `src-nh-nhag-charge` | New Hampshire AG charging announcement; primary record of allegation; A for charge, not guilt | https://www.doj.nh.gov/news-and-media/steven-kramer-charged-voter-suppression-over-ai-generated-president-biden-robocalls |
| `src-nh-ap-acquittal` | Independent report of acquittal and defense account; B | https://apnews.com/article/ai-robocalls-new-hampshire-biden-kramer-e87ef8848a2a6515202535a6e800da07 |
| `src-nh-lwv-case` | Civil-case document index and outcome summary; B / links primary filings | https://www.lwv.org/legal-center/league-women-voters-new-hampshire-v-kramer |
| `src-nh-ap-default` | Independent report of default judgment, $22,500 award, injunction, and refusal to pay; B | https://apnews.com/article/7bb71c5ab962760995469c3b6f787a86 |

---

## 2. `bra-2026-bolsonaro-ai-avatar-flavio-launch`

### Proposed case record

| Field | Value |
|---|---|
| `case_id` | `bra-2026-bolsonaro-ai-avatar-flavio-launch` |
| `case_family_id` | `bra-2026-general-election` |
| Country | Brazil |
| Election/date | 2026 general election: first round 2026-10-04; possible second round 2026-10-25; avatar shown 2026-07-25 |
| Incident | AI avatar of Jair Bolsonaro endorsing Flávio Bolsonaro at the PL convention/campaign launch |
| Court file | **TSE Representation `RP 0601315-97.2026.6.00.0000`** |
| Status | `ongoing` and legally `contested`; cut-off 2026-08-11 |
| Actor / attribution | PL/Flávio Bolsonaro campaign presented the asset at its convention and defends it. Jair Bolsonaro's own knowledge/authorization is denied by his lawyers and remains unresolved. Do not collapse producer/presenter, beneficiary, depicted person, and authorizer into one actor. |
| AI/digital capability | Confirmed synthetic video/voice avatar of a real political leader; explicit AI disclosure at the beginning |
| Influence vector/platform | Large-screen convention presentation followed by online/social and news circulation; verified unique online exposure is unavailable |
| Target | PL convention audience, Bolsonaro supporters, and the broader 2026 electorate; immediate beneficiary is Flávio Bolsonaro's candidacy |
| Mechanism | A synthetic surrogate placed Jair's recognizable image, voice, affect, and first-person endorsement into a campaign event while the real Jair was ineligible and subject to public-communication restrictions |
| Scale/reach | Occurrence before the convention audience is established and broad secondary circulation is observable; no defensible unique-viewer or impression count has been published. |
| Behavioural/electoral effect | None established; voting has not occurred as of the cut-off. No claim about votes, candidate support, turnout, or result is admissible yet. |
| Agency-transfer path | Synthetic embodiment lets the Flávio/PL campaign exercise a form of Jair's personal endorsement and charismatic authority without his physical/public participation. The transferred capacity is control over whose voice appears to make a dynastic succession appeal. This is an **authority-substitution/circumvention** mechanism even if viewers know the image is synthetic; it need not depend on believing the clip is authentic footage. |
| Democratic harm | Potential circumvention of communication restrictions and candidacy rules; synthetic appropriation of a real person's political authority; unequal campaign advantage; uncertainty over whether disclosure cures the prohibited use |
| Response/outcome | Federação Brasil da Esperança filed the TSE representation on 2026-07-26 and sought removal/fines; a related notice went to the STF. Flávio/PL contested liability. As of 2026-08-11 there is **no merits judgment**. TSE president Kássio Nunes Marques called an internal meeting for 2026-08-13; that future meeting is a monitoring event, not an outcome. |
| Evidence classification | Occurrence/AI: confirmed; PL presentation/benefit: confirmed; Jair authorization: unknown/denied; reach: qualitative; legal violation: alleged and pending; behaviour/electoral effect: unknown/not yet measurable |
| Overall confidence | 0.97 occurrence; 0.95 AI/disclosure; 0.90 docket/pending posture; 0.40 on which precise rule a future merits panel will apply; 1.00 that no electoral effect can yet be assessed at the cut-off |

### Exact operative rules

The relevant consolidated instrument is TSE Resolution 23.610/2019 as amended by Resolutions 23.732/2024 and 23.755/2026. Code the rule and the application separately.

| Provision | What the text establishes | Application to this incident as of 2026-08-11 |
|---|---|---|
| **Art. 9-B caput** | Electoral propaganda using synthetic multimedia must disclose, explicitly, prominently, and accessibly, that it was fabricated/manipulated and identify the technology. | The video reportedly disclosed at the outset that it was an AI simulation. That supports compliance with the disclosure duty only; it does not automatically resolve the separate prohibitions below. |
| **Art. 9-B §3** | Chatbots, avatars, and synthetic content used to mediate campaign communication with natural persons are subject to Art. 9-B, with “qualquer simulação de interlocução com a pessoa candidata ou outra pessoa real” prohibited. | The challengers can plausibly invoke this language. Whether a one-way convention video constitutes prohibited simulated “interlocução” is a legal question for the TSE, not an established dataset fact. |
| **Art. 9-B §3-A** (2026 addition) | From 72 hours before through 24 hours after the close of voting, publication/republication/paid boosting of new synthetic content using a candidate's or public person's image/voice/manifestation is barred even when labelled. | **Not the operative timing rule for the 2026-07-25 convention**, which occurred far outside the Oct. 4 voting blackout window. Retain it for prospective monitoring only. |
| **Art. 9-C caput** | Bars fabricated/manipulated electoral propaganda that spreads notoriously false or decontextualized facts capable of harming electoral balance or integrity. | It is not yet adjudicated whether the disclosed avatar's message contained the kind of false/decontextualized fact required by the caput. |
| **Art. 9-C §1** | Prohibits synthetic audio/video used to harm or favor a candidacy—even with authorization—when it digitally creates, substitutes, or alters the image or voice of a living, dead, or fictitious person (“deep fake”). | This is the strongest text for the challenge because the avatar favored Flávio and created/substituted Jair's image/voice. Flávio's defense disputes that the disclosed “digital puppet” is a deepfake or deceptive. The court's construction remains unresolved. |
| **Art. 9-C §2** | Violation of the caput/§1 constitutes abuse of political power and misuse of the means of social communication (`uso indevido dos meios de comunicação social`), potentially producing cancellation of registration/mandate and other liability. | These are potential statutory consequences, **not sanctions already imposed**. |

### Atomic claim ledger

| Claim ID | Type | Atomic claim | Label | Conf. | Supporting sources | Counterevidence / constraint |
|---|---|---|---|---:|---|---|
| `clm-bra26-avatar-001` | occurrence / ai_role | On 2026-07-25, the PL convention launching Flávio Bolsonaro's presidential candidacy showed an AI-generated audiovisual avatar of Jair Bolsonaro. | established_evidence | 0.98 | `src-bra26-reuters`; `src-bra26-petition` | No material dispute that the clip was synthetic or shown. |
| `clm-bra26-avatar-002` | mechanism | The avatar spoke in the first person about imprisonment/silencing and asked the audience to embrace the person chosen to replace him, his son Flávio. | established_evidence | 0.97 | `src-bra26-reuters`; `src-bra26-petition` | The asset announced that it was made with AI; it should not be described as wholly undisclosed authentic footage. |
| `clm-bra26-avatar-003` | attribution | PL/Flávio's campaign presented and defended the avatar and was its immediate electoral beneficiary. | established_evidence | 0.96 | `src-bra26-reuters`; `src-bra26-defense-report` | Public evidence does not yet establish Jair's personal authorization or participation in production. |
| `clm-bra26-avatar-004` | attribution / counterevidence | Jair Bolsonaro's lawyers denied that he authorized or knew of the avatar's production. | established_evidence | 0.95 | `src-bra26-ft`; `src-bra26-reuters` | The evidence establishes the denial, not the underlying absence of authorization. Code authorization `unknown`. |
| `clm-bra26-avatar-005` | legal_status | Federação Brasil da Esperança filed TSE case RP 0601315-97.2026.6.00.0000 on 2026-07-26 alleging premature propaganda and unlawful synthetic-media use. | established_evidence | 0.99 | `src-bra26-petition` | The evidence establishes the filing/allegation and docket, not a violation. |
| `clm-bra26-avatar-006` | legal_status | Art. 9-B caput imposes disclosure; Art. 9-B §3 bars simulated interlocution with a candidate/real person; Art. 9-C §1 bars synthetic audio/video used to favor/harm a candidacy even when authorized. | established_evidence | 0.99 | `src-bra26-rule-consolidated` | The legal text is established; its application to this disclosed one-way avatar is unresolved. |
| `clm-bra26-avatar-007` | legal_status / counterevidence | The video's up-front AI disclosure supports compliance with Art. 9-B's labelling duty but does not by itself answer Art. 9-B §3 or Art. 9-C §1. | strong_inference | 0.93 | `src-bra26-reuters`; `src-bra26-rule-consolidated` | Defense argues disclosure and lack of deception remove the asset from “deep fake”; no merits court has accepted or rejected that reading. |
| `clm-bra26-avatar-008` | counterevidence | Art. 9-B §3-A's 72-hour pre-election/24-hour post-election blackout does not apply by timing to the July 25 convention. | established_evidence | 0.99 | `src-bra26-rule-2026`; `src-bra26-tse-explainer` | Other provisions may still apply. |
| `clm-bra26-avatar-009` | reach | The clip reached the convention audience and circulated online/news, but no reliable unique-viewer or impression figure is available. | established_evidence | 0.92 | `src-bra26-reuters`; `src-bra26-petition` | The claim is deliberately limited to qualitative reach and absence of a defensible count; article pickup, repost counts, and convention attendance should not be summed as unique exposure. |
| `clm-bra26-avatar-010` | behavioural_effect | No evidence currently measures whether seeing the avatar changed attitudes, candidate preference, participation, or sharing behavior. | open_question | 0.99 | `src-bra26-reuters` | Persuasive potential alleged in the petition is not an observed behavioural effect. |
| `clm-bra26-avatar-011` | electoral_effect | The October 2026 election has not occurred; the avatar cannot be coded as having changed its outcome. | established_evidence | 1.00 | `src-bra26-tse-explainer` | Any outcome language before the vote is prospective speculation. |
| `clm-bra26-avatar-012` | agency_transfer | The disclosed avatar functions as a synthetic political surrogate: it enables the beneficiary campaign to deploy Jair's recognizable persona and endorsement despite his ineligibility and public-communication restrictions. | strong_inference | 0.91 | `src-bra26-reuters`; `src-bra26-ft` | Whether this legally “circumvented” a specific court restriction depends in part on authorization and the STF/TSE rulings, which remain unresolved. |
| `clm-bra26-avatar-013` | institutional_outcome | As of 2026-08-11 the TSE has not issued a merits decision; an internal meeting was scheduled for Aug. 13. | established_evidence | 0.94 | `src-bra26-pending` | Do not backfill the scheduled meeting as though it had happened. Reverify after Aug. 13. |

### Sources

| Source ID | Role / quality | Exact URL |
|---|---|---|
| `src-bra26-petition` | FE Brasil representation, dated 2026-07-26; primary record of claims/docket; A for filing | https://static.congressoemfoco.com.br/2026/07/27/attachment/2026/07/27/cd5726_tse_fe_brasil_ia_jair_convencao_pl_26_jul__2026.pdf |
| `src-bra26-rule-consolidated` | Consolidated TSE Resolution 23.610, arts. 9-B/9-C; primary law; A | https://www.tse.jus.br/legislacao/compilada/res/2019/resolucao-no-23-610-de-18-de-dezembro-de-2019 |
| `src-bra26-rule-2026` | TSE Resolution 23.755/2026 amendments; primary law; A | https://www.tse.jus.br/legislacao/compilada/res/2026/resolucao-no-23-755-de-2-de-marco-de-2026 |
| `src-bra26-tse-explainer` | TSE 2026 rule/election-date explainer; primary authority; A | https://www.tse.jus.br/comunicacao/noticias/2026/Abril/por-dentro-das-eleicoes-conheca-as-regras-sobre-uso-de-ia-na-campanha-eleitoral-de-2026 |
| `src-bra26-reuters` | Independent occurrence, message, disclosure, filing, defense, and pending-status report; B | https://www.reuters.com/business/media-telecom/brazils-bolsonaro-is-barred-this-years-election-can-his-ai-avatar-campaign-2026-07-28/ |
| `src-bra26-ft` | Independent report of counsel's denial of authorization and legal challenge; B | https://www.ft.com/content/e18c380f-c25d-4e91-8135-d9c2424ab720 |
| `src-bra26-defense-report` | Defense-position reporting; B | https://www1.folha.uol.com.br/poder/2026/07/flavio-diz-ao-tse-que-video-de-ia-com-bolsonaro-e-legal-e-nao-engana-o-eleitorado.shtml |
| `src-bra26-pending` | Same-day report that Nunes Marques scheduled internal TSE discussion for 2026-08-13; B, ongoing monitoring | https://www.poder360.com.br/poder-justica/kassio-marca-reuniao-com-ministros-do-tse-para-definir-ia-de-bolsonaro/ |

---

## 3. `usa-2026-nrsc-talarico-synthetic-ad`

### Proposed case record

| Field | Value |
|---|---|
| `case_id` | `usa-2026-nrsc-talarico-synthetic-ad` |
| `case_family_id` | `usa-2026-midterm-election` |
| Country | United States (Texas) |
| Election/date | Texas U.S. Senate general election, 2026-11-03; NRSC ad released 2026-03-11 |
| Incident | NRSC “James Talarico Reads His Most Memorable Tweets” synthetic video |
| Status | `ongoing`; observed campaign tactic in an unfinished election |
| Actor / attribution | National Republican Senatorial Committee: `admitted`; it published the ad, labelled the speaker “AI James Talarico” in its own transcript, and defended the use |
| AI/digital capability | Confirmed high-realism synthetic face and cloned/synthetic voice; first-person embodiment of a living candidate; small on-screen “AI generated” disclosure |
| Influence vector/platform | NRSC web release, YouTube and X/social distribution, followed by earned-media circulation |
| Target | Texas voters and national audiences evaluating Democratic Senate candidate James Talarico |
| Mechanism | The ad converts older written statements/posts into an apparent current audiovisual performance by Talarico and inserts reaction lines/affect that were not part of the underlying posts (“So true, so true”; “Should have been a preacher”). |
| Scale/reach | Public release is established. A snapshot of the NRSC YouTube upload showed roughly 3.7K views on 2026-08-11, but the figure is dynamic, platform-specific, and not a total campaign reach estimate. X/cross-platform impressions and unique exposure are unavailable. |
| Behavioural/electoral effect | None measured. The election is in the future, and no ad-specific persuasion, belief, sharing, or vote-choice study is available. Reuters' general expert/study discussion cannot be assigned as this ad's effect. |
| Agency-transfer path | Synthetic embodiment transfers control over Talarico's apparent present-tense voice, facial affect, and self-evaluation to an opposing party committee. Unlike a screenshot of real posts, it makes the target appear to perform and endorse the framing. |
| Democratic harm | Candidate-identity appropriation; confusion between authentic quote, synthetic enactment, and invented commentary; normalization of opponent replicas; potential erosion of audiovisual evidence. Mitigated—but not eliminated—by an AI label, explicit sponsor, and authentic underlying posts. |
| Response/outcome | Talarico campaign called it misleading; NRSC defended it as presenting his own words. No reported removal or enforcement. Reuters reported no federal regulation constraining this use and a patchwork of state laws. |
| Evidence classification | Occurrence/AI/actor: confirmed/admitted; disclosure: confirmed but salience disputed; reach: limited platform snapshot + unknown cross-platform; deception/significance: contested; behavioural/electoral effect: unknown/not yet measurable |
| Overall confidence | 0.99 occurrence/actor/AI; 0.95 composition of transcript; 0.70 for watermark-salience assessment; 0.20 for audience deception; 1.00 that no electoral effect can yet be assessed at the cut-off |

### Atomic claim ledger

| Claim ID | Type | Atomic claim | Label | Conf. | Supporting sources | Counterevidence / constraint |
|---|---|---|---|---:|---|---|
| `clm-us26-tal-001` | occurrence / attribution | On 2026-03-11 the NRSC released a digital ad depicting an AI-generated James Talarico reading earlier statements/posts. | established_evidence | 0.99 | `src-us26-nrsc`; `src-us26-reuters` | Actor attribution is not inferential: the NRSC published and transcribed it. |
| `clm-us26-tal-002` | ai_role | The ad used a realistic synthetic face and voice; the words “AI generated” appeared in small type at lower right. | established_evidence | 0.98 | `src-us26-reuters`; `src-us26-berkeley` | The existence of a label is established. “Easy to miss” is Reuters' assessment, not a measured disclosure-recognition rate. |
| `clm-us26-tal-003` | mechanism | Much of the substantive attack used genuine older Talarico posts, but the synthetic first-person performance included added reaction lines not contained in those quoted posts. | established_evidence | 0.97 | `src-us26-nrsc`; `src-us26-reuters` | Do not describe every spoken proposition as fabricated; distinguish authentic underlying text from invented performance/asides. |
| `clm-us26-tal-004` | counterevidence | The ad identified its sponsor, carried an AI disclosure, and the NRSC publicly described it as AI; these facts reduce covert-source deception. | established_evidence | 0.98 | `src-us26-nrsc`; `src-us26-reuters` | Disclosure does not turn the inserted affect/phrasing into Talarico's speech or show that viewers noticed/understood the label. |
| `clm-us26-tal-005` | reach | The NRSC upload was publicly viewable; a 2026-08-11 YouTube snapshot displayed about 3.7K views. | established_evidence | 0.88 | `src-us26-youtube` | Code the case-level `reach_status` as `platform_reported`. This is a dynamic, non-unique, one-platform count; do not combine it with article readership or X reposts. |
| `clm-us26-tal-006` | behavioural_effect | No ad-specific evidence establishes whether viewers believed it was authentic footage, updated their view of Talarico, or shared it because of the synthetic format. | open_question | 0.99 | `src-us26-reuters`; `src-us26-berkeley` | General deepfake-detection research is contextual, not this incident's effect estimate. |
| `clm-us26-tal-007` | electoral_effect | As of 2026-08-11 the Nov. 3 election has not occurred; no electoral effect can be recorded. | established_evidence | 1.00 | `src-us26-reuters` | Any “effective ad” language from strategists is not a vote-change estimate. |
| `clm-us26-tal-008` | agency_transfer | By making the opponent appear to speak and react on camera, the sponsor controlled not merely message selection but the target's simulated voice, face, timing, and affect. | strong_inference | 0.94 | `src-us26-nrsc`; `src-us26-reuters`; `src-us26-berkeley` | The underlying posts constrain how much substantive policy content was invented. |
| `clm-us26-tal-009` | response / legal_status | The NRSC defended the ad; the Talarico campaign called it misleading; no removal/enforcement was reported, and Reuters described no federal rule specifically constraining such political AI messaging. | established_evidence | 0.96 | `src-us26-reuters` | State-law coverage varies. Do not state that all US deepfake political advertising is lawful. |

### Sources

| Source ID | Role / quality | Exact URL |
|---|---|---|
| `src-us26-nrsc` | Sponsor's release and full transcript; primary actor record; A for what NRSC published, C for evaluative claims | https://www.nrsc.org/press-releases/new-ad-james-talarico-reads-his-most-memorable-tweets-2026-03-11/ |
| `src-us26-youtube` | Sponsor's platform upload; primary artifact/dynamic platform metric | https://www.youtube.com/watch?v=qIDhhVah4xE |
| `src-us26-x` | Sponsor's X distribution; primary artifact | https://x.com/NRSC/status/2031871266609664396?lang=en |
| `src-us26-reuters` | Independent occurrence, disclosure, campaign response, and legal-context report; B | https://www.reuters.com/business/media-telecom/ai-deepfakes-blur-reality-2026-us-midterm-campaigns-2026-03-28/ |
| `src-us26-berkeley` | UC Berkeley summary of Hany Farid's forensic assessment; B | https://www.ischool.berkeley.edu/news/presscoverage/2026/hany-farid-delves-newly-released-deepfakes-james-talarico |
| `src-us26-texas-law` | Texas/federal-law contextual reporting; B | https://www.houstonchronicle.com/politics/texas/article/ai-ads-texas-21957736.php |

---

## Immediate coding cautions

1. **New Hampshire:** keep the final FCC finding, criminal acquittal, and civil default judgment as three distinct legal propositions. “Acquitted” does not negate admitted orchestration; “FCC intent finding” is not a criminal conviction.
2. **Brazil:** the case number is `RP 0601315-97.2026.6.00.0000`. Code the FE Brasil assertions as allegations. The disclosure is counterevidence, not a complete defense; the TSE merits question is open. Art. 9-B §3-A's election-period blackout is temporally inapplicable to July 25.
3. **US 2026:** the Talarico artifact is an observed campaign ad, not evidence that deepfakes will determine the midterms. Preserve the distinction between real underlying posts, synthetic embodiment, and invented asides.
4. **All three:** do not infer belief, turnout, vote choice, or election outcome from call attempts, views, media coverage, or legal alarm.